stream the reply: first ink ~1.4s after commit (was ~4s+)

- oracle: gpt-5.4-mini + --no-tools (reads handwriting as well as gpt-5.5,
  streams ~3x sooner, measured via pi rpc bench); deliver each completed
  sentence as it streams, drop the sender on agent_end (disconnect = done).
- extract_assistant_text: read only the FIRST assistant section of an event —
  message_update lines carry the partial twice (assistantMessageEvent.partial
  + top-level message) and would double every streamed chunk; agent_end's
  messages array is no longer used for text (it concatenates the whole
  session's messages from turn 2 on).
- main: ask the oracle when the ink commits, so the model streams during the
  drink dissolve; streamed continuation chunks splice into the running write
  animation on new lines (WritePlan.next_y).
- ink: downscale to <=800px long side (adaptive, min 2x), fast PNG deflate.
- log time from ask to first delivered chunk.

Measured on device, full takeover, 2-turn session: first chunk +1461ms /
+1319ms after ask; replies clean and in character on both turns.